【摘要】 针对现有算法在交通场景下对于车辆行人等目标检测精度不足、检测速度较慢的问题，在YOLOv4中使用MobileNetV3结构替换YOLOv4的主干特征提取网络，通过MobileNetV3中的深度可分离卷积大幅降低主干网络参数量，并将SPP结构使用跨阶段局部网络结构进行改造，消除信息冗余现象。最后在网络结构中加入注意力模块CBAM，提升检测精度。通过在UA-DETRAC数据集上进行实验，该方法平均检测精度达到92.6%，相较于YOLOv4，在模型大小下降24%的情况下，平均检测精度提升1.7%，平均漏检率降低11%，召回率提升1.6%，结果显示该算法在优化交通场景目标检测上具有可行性。更多还原

【Abstract】 Aiming at the problem that the existing algorithm has insufficient detection accuracy and slow detection speed for targets such as vehicle pedestrians in traffic scenarios, the backbone feature extraction network of YOLOv4 is replaced by MobileNetV3 structure in YOLOv4, and the amount of trunk network parameters is greatly reduced through deep separable convolution in MobileNetV3, and then the SPP structure is transformed by using the structure of the cross-stage local network to eliminate information redundancy. Finally, the attention module CBAM is added to the network structure to improve the detection accuracy. Through experiments on the UA-DETRAC dataset, the average detection accuracy of the method reached 92.6% with 24% reduction of model size, compared with YOLOv4, the average detection accuracy was increased by 1.7%, the average missed detection rate was reduced by 11%, and the recall rate was increased by 1.6%. The results show that the algorithm is feasible in optimizing traffic scene target detection.更多还原